Research map / Label-efficient and robust learning / Domain adaptation and generalization
Towards Unsupervised Domain Generalization for Face Anti-Spoofing
Yuchen Liu, Yabo Chen, Mengran Gou, Chun-Ting Huang, Yaoming Wang, Wenrui Dai, Hongkai Xiong
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Split-Rotation-Merge module (ablation) | UDG-FAS · FAS I&C&M to O | HTER (%): 17.93 -> 12.18 | Improved |
| add in-domain nearest-neighbor positives (ablation) | UDG-FAS · FAS I&C&M to O | HTER (%): 14.31 -> 12.18 | Improved |
| add cross-domain nearest-neighbor positives (ablation) | UDG-FAS · FAS I&C&M to O | HTER (%): 16.08 -> 12.18 | Improved |
| add domain-specific normalization before cross-domain NN search (ablation) | UDG-FAS · FAS I&C&M to O | HTER (%): 13.94 -> 12.18 | Improved |
| add merged local features to cross-domain NN distance (ablation) | UDG-FAS · FAS I&C&M to O | HTER (%): 12.97 -> 12.18 | Improved |
| use SRM instead of patch shuffle augmentation (ablation) | UDG-FAS · FAS I&C&M to O | HTER (%): 15.67 -> 12.18 | Improved |
| increase number of nearest neighbors k from 1 to 4 (ablation) | UDG-FAS · FAS I&C&M to O | HTER (%): 12.18 -> 14.24 | Hurt |
| add unlabeled CelebA-Spoof data (200k images) to pretraining (ablation) | UDG-FAS · FAS I&C&M to O | HTER (%): 12.18 -> 9.06 | Improved |
| use UDG-FAS unsupervised pretraining instead of ImageNet init | ResNet-18 FAS · UDG-Protocol-1 | average AUC: +9.43% average AUC | Improved |
| use UDG-FAS instead of transformer SSL (MAE) for pretraining | FAS pretraining · FAS I&C&M to O | HTER (%): 15.28 -> 12.18 | Improved |
| combine UDG-FAS pretraining with SSDG finetuning | SSDG-R · UDG-Protocol-1 | average HTER: 11.29% -> 8.15% | Improved |
| add UDG-FAS to SimSiam-style contrastive learning vs generic methods (MoCo, SimCLR, BYOL) | unsupervised pretraining · UDG-Protocol-1 FAS | HTER (%): I&C&M to O: 26.67 (SimCLR V2) -> 12.18 | Improved |
Closest papers by meaning
- Rethinking Domain Generalization for Face Anti-Spoofing: Separability and Alignment (CVPR 2023 · similarity 0.94)
- Single-Side Domain Generalization for Face Anti-Spoofing (CVPR 2020 · similarity 0.94)
- Gradient Alignment for Cross-Domain Face Anti-Spoofing (CVPR 2024 · similarity 0.93)
- Test-Time Domain Generalization for Face Anti-Spoofing (CVPR 2024 · similarity 0.93)
- Rethinking Generalizable Face Anti-spoofing via Hierarchical Prototype-guided Distribution Refinement in Hyperbolic Space (CVPR 2024 · similarity 0.90)
- CFPL-FAS: Class Free Prompt Learning for Generalizable Face Anti-spoofing (CVPR 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
